“A system prompt for Claude was designed to ensure symmetrical treatment of political topics, preventing the model from refusing tasks related to one political leaning while accepting equivalent tasks for another.”